Performance and optimizing bandwidth

[dabreegster]

Currently the raw OS polygons are stored in a flatgeobuffer, with a style code for debugging. There are many network requests to fetch an area, and the API call can easily take more than 30s for relatively small routes.

• [x] Pre-filter the OS polygons for just the ones we want, and omit the style code property
• [x] Dissolve adjacent polygons. We don't need to see individual buildings; one polygon covering a block is less bytes transferred and geometry checks
• [ ] Try https://github.com/michaelkirk/geomedea instead of fgb. It now has WASM support!

[dabreegster]

Seeing benefits for geomedea in https://github.com/a-b-street/15m/pull/8. The use case in this repo is much simpler -- just going to read polygons with no properties.

[dabreegster]

The polygon dissolving can be over-eager; the huge shape below is one big polygon! This really slows down perpendicular checks, because we look at loads of polygon edges that're irrelevant and the rtree can't filter out.

[dabreegster]

I don't think there's an easy way to limit total area with mapshaper dissolve. Some other possible sources for dissolve logic:

• https://github.com/emilyselwood/polygonical/blob/f03f37cf016cea873fd5fbdf3d72915795f97817/src/polygon.rs#L224
• https://gitlab.com/bronsonbdevost/rust-geo-repair-polygon/-/blob/master/src/join.rs?ref_type=heads

[dabreegster]

512 perpendiculars, with a total of 110537699 line hit checks, taking over 12s total

Going to try georust booleaps, and if they have problems, the new spade versions